Research map / Large language models / Reasoning and chain of thought
TIV: Thought Injection via Vectors for Efficient Reasoning in Large Reasoning Models
Yi Cao, Weijie Shi, Wei-Jie Xu, Yucheng Shen, Yue Cui, Hanghui Guo, Shimin Di, Ziyi Liu, Jiaming Li, Alexander Zhou, Jia Zhu, Jiajie Xu
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes TIV, a framework that compresses token-level reasoning traces into thought vectors that are injected to make large reasoning models reason more efficiently and reduce overthinking.
Problem
Large reasoning models often overthink, generating excessive tokens that add little to final accuracy while increasing inference cost.
Approach
Compresses token-level reasoning into vector representations ("thought vectors") and injects them back into the model to guide reasoning with fewer tokens, rather than relying on explicit token-level chains.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add vector injection to the compression stage (ablation) | TIV (R1-1.5B) · AIME 2024 | AES: 0.03 -> 0.78 | Improved |
| add vector injection to the compression stage (ablation) | TIV (R1-1.5B) · GSM8K | AES: 0.41 -> 0.65 | Improved |
| add length penalty to the reward (ablation) | TIV (R1-1.5B) · MATH 500 | avg tokens: 2811.1 -> 1283.3 | Improved |
| add length penalty to the reward (ablation) | TIV (R1-1.5B) · AIME 2024 | AES: 0.63 -> 0.78 AES, accuracy 30.8 unchanged, length 9021.0 -> 6605.4 | Mixed |
| add Stage I reasoning calibration (ablation) | TIV (R1-1.5B) · AIME 2024 | accuracy (%): 24.2 -> 30.8 | Improved |
| inject vectors after MLP as well as MHA (ablation) | TIV (R1-1.5B) · AIME 2024 | accuracy (%): 30.8 -> 32.5 | Improved |
| increase number of injected vectors beyond one (ablation) | TIV · training reward and length | reward | No clear effect |
| increase rollouts per prompt beyond 5 (ablation) | TIV · training reward and length | reward | No clear effect |
| apply TIV to R1-7B reasoning | DeepSeek-R1-Distill-Qwen-7B · AIME 2024 | accuracy (%): 52.9 -> 48.5 | Hurt |
| apply TIV to R1-7B reasoning | DeepSeek-R1-Distill-Qwen-7B · MATH 500 | accuracy (%): 91.2 -> 88.6 | Hurt |
| apply TIV to R1-7B reasoning | DeepSeek-R1-Distill-Qwen-7B · GSM8K | avg tokens: 725.7 -> 110.2 | Improved |
| apply TIV to R1-1.5B on out-of-domain MMLU | R1-1.5B · MMLU | accuracy (%): 29.1 -> 33.7 | Improved |
Closest papers by meaning
- Thinking on the Fly: Test-Time Reasoning Enhancement via Latent Thought Policy Optimization (ICLR 2026 · similarity 0.89)
- How Far Are We from Optimal Reasoning Efficiency? (NeurIPS 2025 · similarity 0.89)
- On Reasoning Strength Planning in Large Reasoning Models (NeurIPS 2025 · similarity 0.89)
- Training Large Reasoning Models Efficiently via Progressive Thought Encoding (ICLR 2026 · similarity 0.89)
- Don’t Think Longer, Think Wisely: Optimizing Thinking Dynamics for Large Reasoning Models (NeurIPS 2025 · similarity 0.88)
- Efficient Reasoning via Constrained Optimization in Latent Space (NeurIPS 2026 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
